Agents

The Agent is Spark's agentic layer: something you can talk to that can also take real action across your workspace. It's one of the two sides of Spark, and it only works because of the other side, the context around it.

Agents are teammates

The simplest way to think about an agent is as a colleague. A workspace is less a piece of software you operate and more a team you work within: people and agents sit side by side, and you hand an agent a piece of work much as you would a coworker.

You can build more than one, each with its own role, and they cooperate with your team rather than replace it. One agent might watch for incoming files, another might answer questions about a dataset, while people handle the judgment calls that need a human. Giving an agent a clear name, description, and set of instructions is how you define that role.

Context is what makes it useful

An agent bolted onto a product with no knowledge of your data is just a chatbot. Spark's Agent is different because it runs inside your workspace and can already see your data objects, Spaces, canvases, tasks, files, and workflows. You don't paste data into a prompt; the Agent reads the real thing and acts on it.

That's the core idea to hold onto: an agent is only as good as the context it can see, and Spark's job is to give it a rich, structured one.

What shapes an agent

An agent is more than a model. A few things define what a given agent is and can do:

• A name and description give it an identity, the way a role on a team has a title and a remit.
• Its instructions are its standing brief: what it's for and how it should behave, much like a job description.
• Tools are the specific capabilities it can invoke to take an action.
• Skills are packaged instructions that shape how it handles a particular kind of task.
• Connectors let it reach external systems, including over MCP (Model Context Protocol).
• A vault holds secrets, so the Agent can authenticate to those systems without ever exposing credentials.

Configuring an agent is mostly a matter of writing its instructions and deciding which of these it should have for the job you have in mind.

Working with an agent

You interact with an agent in a chat, where you tag it with @ and give it work. Because it carries your workspace context, it can summarize records, draft output, find work that needs attention, or carry out multi-step actions on your behalf.

Agents and workflows

The Agent handles the parts of work that need reasoning and judgment, where the next step isn't fixed in advance. Workflows handle the parts you want to run the same way every time. They're complementary: reach for a workflow when you want a guarantee, and the Agent when you want judgment.

The Agent section covers how to create, configure, and chat with agents.